On Fri 22-09-23 16:00:30, Roman Gushchin wrote:
> On Wed, Sep 20, 2023 at 03:47:37PM +0200, Michal Hocko wrote:
> > On Wed 20-09-23 15:25:23, Jeremi Piotrowski wrote:
> > > On 9/20/2023 1:07 PM, Michal Hocko wrote:
> > [...]
> > > > I mean, normally I would be just fine reverting this API change because
> > > > it is disruptive but the only way to have the file available and not
> > > > break somebody is to revert 58056f77502f ("memcg, kmem: further
> > > > deprecate kmem.limit_in_bytes") as well. Or to ignore any value written
> > > > there but that sounds rather dubious. Although one could argue this
> > > > would mimic nokmem kernel option.
> > > >
> > >
> > > I just want to make sure we don't introduce yet another new behavior in this legacy
> > > system. I have not seen breakage due to 58056f77502f. Mimicing nokmem sounds good but
> > > does this mean "don't enforce limits" (that should be fine) or "ignore writes to the limit"
> > > (=don't event store the written limit). The latter might have unintended consequences.
> >
> > Yes it would mean that the limit is never enforced. Bad as it is the
> > thing is that the hard limit on kernel memory is broken by design and
> > unfixable. This causes all sorts of unexpected kernel allocation
> > failures that this is simply unsafe to use.
> >
> > All that being said I can see the following options
> > 1) keep the current upstream status and not export the file
> > 2) revert both 58056f77502f and 86327e8eb94 and make it clear
> > that kmem.limit_in_bytes is unsupported so failures or misbehavior
> > as a result of the limit being hit are likely not going to be
> > investigated or fixed.
> > 3) reverting like in 2) but never inforce the limit (so basically nokmem
> > semantic)
>
> Since it's a part of cgroup v1 interface, which is in a frozen state as a whole,
> and there is no significant (performance, code complexity) benefit of
> additionally deprecating kmem.limit_in_bytes, I vote for 2).
> 1) is also an option.
We have a stronger agrement over 3)
http://lkml.kernel.org/r/ZRE5VJozPZt9bRPy@xxxxxxxxxxxxxx. Please speak
up if you disagree.
--
Michal Hocko
SUSE Labs
